DecisionRank trains decision model rerankers
Original title: DecisionRank: On-Policy Distillation for Decision-Model Rerankers
DecisionRank achieves SOTA on TREC DL and BEIR with a 1.7B model that surpasses its 27B generative teacher.
DecisionRank uses item-space on-policy distillation where a generative teacher provides feedback on sorted document scores from the student. Combined with offline supervision it sets new state-of-the-art on TREC DL, BEIR and BRIGHT. The 1.7B model beats the 27B teacher on TREC DL19-23 and the 8B version runs 24.5 times faster on TREC DL.
- Item-space OPD consistently outperforms alternative objectives
- 1.7B model surpasses 27B teacher on TREC DL19-23
- 8B model runs 24.5x faster than listwise counterpart
- SOTA across general and reasoning-intensive retrieval benchmarks
- AI models
- DecisionRank27B generative teacher
From the original article
1 min read at huggingface.co
Decision models target structured tasks by directly predicting choices or scores, avoiding autoregressive decoding latency. However, transferring a generative LLM's ranking knowledge to a decision model remains challenging: the teacher expresses preferences through a ranked list of document ID sequences, while the student determines the ranking using scores from the first-token position.
By Yanran Wu, Dian Yu, Chongyang Bai · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.